LocalEvalService.evaluate adds rubrics to the caller's inference results, so scoring them again raises #7874

🔴 Required Information

Describe the Bug:
LocalEvalService.evaluate modifies the inference results it is given. If you score the same results twice, the second call raises and aborts the whole run.

_evaluate_single_inference_result copies the eval case's rubrics, and each expected invocation's rubrics, onto inference_result.inferences. Those are the caller's own Invocation objects. A second evaluate over them finds the rubric ids already present, and _add_rubrics_to_invocation raises ValueError: Rubric with rubric_id '...' already exists.

That happens outside the per-metric error handling, so nothing is scored or saved. It happens for any metric, even ones that ignore rubrics such as response_match_score. Results serialized after a first evaluate carry the rubrics too, so they fail the same way.

This breaks the flow in the eval service guide's "Score a run again without rerunning it" section: keep a run and score it again later against a stricter threshold or an extra metric.

Steps to Reproduce:

1. Add an eval case with a case-level rubric, then run perform_inference on it.
2. Call evaluate on the inference results, with any metric.
3. Call evaluate on the same results again, e.g. with a stricter threshold.

Expected Behavior:
Both evaluate calls score the run, and the caller's inference results come back unchanged.

Observed Behavior:

1st evaluate: PASSED | caller's inference now carries rubrics: ['names_device']
2nd evaluate (same objects) raised: ValueError Rubric with rubric_id 'names_device' already exists.
evaluate on results serialized BEFORE first evaluate: PASSED
evaluate on results serialized AFTER first evaluate raised: ValueError Rubric with rubric_id 'names_device' already exists.

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at e9e8f61)
• Desktop OS: macOS
• Python Version (python -V): 3.11.16

Model Information:

• Are you using LiteLLM: No
• Which model is being used: any (the repro uses a fake BaseLlm; scoring with response_match_score needs no model)

🟡 Optional Information

Minimal Reproduction Code:

mgr = InMemoryEvalSetsManager()
mgr.create_eval_set(app_name="app", eval_set_id="s1")
mgr.add_eval_case(app_name="app", eval_set_id="s1", eval_case=EvalCase(
    eval_id="c1",
    conversation=[Invocation(
        user_content=types.Content(role="user", parts=[types.Part(text="Turn off the bedroom light.")]),
        final_response=types.Content(role="model", parts=[types.Part(text="The bedroom light is off.")]))],
    rubrics=[Rubric(rubric_id="names_device", rubric_content=RubricContent(text_property="Names the device."))]))
svc = LocalEvalService(root_agent=agent, eval_sets_manager=mgr)

infs = [r async for r in svc.perform_inference(
    InferenceRequest(app_name="app", eval_set_id="s1", inference_config=InferenceConfig()))]
loose = EvaluateConfig(eval_metrics=[EvalMetric(metric_name="response_match_score", threshold=0.5)])
strict = EvaluateConfig(eval_metrics=[EvalMetric(metric_name="response_match_score", threshold=0.9)])

[r async for r in svc.evaluate(EvaluateRequest(inference_results=infs, evaluate_config=loose))]
[r async for r in svc.evaluate(EvaluateRequest(inference_results=infs, evaluate_config=strict))]  # ValueError

How often has this issue occurred?:

• Always (100%), when the eval case or its expected invocations have rubrics